y-kishinami / TGCP

Target-Guided Open-Domain Conversation Planning

This repository contains the code to construct the test set we developed in:

Yosuke Kishinami, Reina Akama, Shiki Sato, Ryoko Tokuhisa, Jun Suzuki and Kentaro Inui. Target-Guided Open-Domain Conversation Planning.

Requirements

nltk==3.4

Test set construction

1. Download start_corpus.pkl and target_set.pkl from the ConvAI2 dataset published on https://github.com/zhongpeixiang/CKC

2. Run the following command

python construct_tgcp_dataset.py --start_utter data/convai2/start_corpus.pkl --target_word data/convai2/target_set.pkl --id id.txt --output {output file}

Evaluation of a generated conversation plan

You can evaluate the conversation plan in the TGCP task by run the following command:

python evaluate.py --input data/result.jsonl

Note: result.jsonl should be in the following jsonl format.

conversation_plan should separate utterances with [SEP].

{"target_word": "book", "conversation_plan": "Hi, What do you do for living?[SEP]I work as an engineer.[SEP]That sounds nice. How do you learn coding?[SEP]I read and learn from technical books."}
{"target_word": "dog", "conversation_plan": "..."}
...

Citation

If you use this If you use anything in this repository, please cite:

Yosuke Kishinami, Reina Akama, Shiki Sato, Ryoko Tokuhisa, Jun Suzuki and Kentaro Inui. Target-Guided Open-Domain Conversation Planning. Proceedings of the 29th International Conference on Computational Linguistics, Oct. 2022.

@InProceedings{kishinami2022tgcp,
  title={Target-Guided Open-Domain Conversation Planning}
  author={Yosuke Kishinami and Reina Akama and Shiki Sato and Ryoko Tokuhisa and Jun Suzuki and Kentaro Inui}
  booktitle={Proceedings of the 29th International Conference on Computational Linguistics}
  year={2022}
}